AI Product Manager — I ship agents that act, and the guardrails that make that safe.
Dubai · pavan.blog · LinkedIn
Most AI demos answer. The hard part is letting a system act — on real money, real customers, real infrastructure — without a human checking every call. That's the problem I work on.
I run a fleet of production agents across live venues, and the discipline matters more than any single model:
- 60+ candidate strategies evaluated → 4 promoted. Over 90% were killed by their own test batteries before touching capital.
- Every candidate declares its promotion criteria before seeing results, then runs shadow → paper → live. No overrides; promotions are gate-driven and audit-logged.
- Gate on irreversible, not uncertain. Confidence is the wrong axis — a low-confidence reversible action is safer than a high-confidence irreversible one.
- Author of RFC #7218, on preventing catastrophic actions in long-running autonomous agents.
| Repo | What it is |
|---|---|
| alphagrid-orchestrator | The production pattern for safely deploying autonomous systems that act on real resources — risk guardian, kill switch, staged-promotion gates. Start here. |
| autonomous-execution-agents | The multi-strategy execution fleet the pattern above governs. |
| content-research-agent | Multi-step Claude agents running content research end to end. Said no to RAG — a structured prompt was enough. |
| dubai-re-intelligence | Flask + Pandas pipeline turning raw Dubai Land Department data into decision intelligence. |
| pavan-blog | My site, including a streaming digital clone built on the Claude API. |
Live systems and production adapters stay private; the reusable patterns are public.
- Glasshouse — a transparent research desk. The product is verifiability: strategies are killed by their own falsification batteries, and the copy-service model is non-custodial.
- Insight Bay — AI automation for UAE SMEs. First paying customer live.
I write at pavan.blog about what actually breaks in production AI — including a case study on shipping agents that act without a human, and what a year of running production agents taught me about reliability.
Daily stack: Claude Code (hooks, MCP servers, slash commands), Anthropic Agent SDK, tool-use / function-calling, LLM evals. Came up as a software engineer — I build the prototypes myself.